dac-research/longbench_synthetic_v3

LongBench Synthetic V3 Dataset statistics Per-subset stats over uploaded samples. Unique ctx counts distinct context strings, while the context-length buckets count samples/rows. Token counts use Qwen/Qwen3-14B. Pool Subset Unique ctx Sample rows <8K 8-16K 16-32K >32K Median tok p90 tok Max tok eval 2wikimqa 197 197 139 54 4 0 6,627 13,271 16,982 eval gov_report 200 200 89 84 24 3 8,902 18,212 52,521 eval musique 200 200 3 46 151 0 16,733 17,231 17… See the full description on the dataset page: https://huggingface.co/datasets/dac-research/longbench_synthetic_v3.